turntax

Coding agents re-send their instruction files with every turn. turntax measures that fixed per-turn cost for a repository, for every agent that reads it, and shows where to cut it.

npx turntax

Try it in the browser, with nothing uploaded: https://torutdev.github.io/turntax/ — drop a folder or enter a public GitHub repository. Links like https://torutdev.github.io/turntax/?repo=owner/name open straight to that repository's report.

State of agent instruction files: what turntax finds in 100 popular repositories, a new edition every month with pinned commits and the script to reproduce it (scripts/state-of.mjs).

Set up the GitHub Action and a config in one command:

npx turntax init

Supported agents

Each agent's documented loading rules decide what is sent on every turn, what loads only when the agent works in a folder or touches a matching file, and what is ignored.

Where an agent reads the same skill name from two folders, it is counted once, in the order the agent documents (or an assumed order, noted for that agent).

Agents that load exactly the same files share one row in the report (files loaded only as a /command, which cost nothing per turn, do not split a row):

always  o200k/turn  on-demand  model                tokens/turn  counted by                   100 turns  cached   agents
4       272         53         (no selected model)                                                                Claude Code
1       63          0          GPT-6.1 Sol          63           estimate (o200k_base proxy)  $0.01      $0.0006  OpenAI Codex, Roo Code, Kiro, JetBrains Junie
1       17          0          (no selected model)                                                                Gemini CLI
5       311         59         GPT-6.1 Sol          311          estimate (o200k_base proxy)  $0.06      $0.0031  Cursor
4       307         55         GPT-6.1 Sol          307          estimate (o200k_base proxy)  $0.06      $0.0031  GitHub Copilot
2       75          0          GPT-6.1 Sol          75           estimate (o200k_base proxy)  $0.01      $0.0008  Windsurf Cascade
2       21          32         GPT-6.1 Sol          21           estimate (o200k_base proxy)  $0.0042    $0.0002  Cline
1       11          0          GPT-6.1 Sol          11           estimate (o200k_base proxy)  $0.0022    $0.0001  Zed
3       84          32         GPT-6.1 Sol          84           estimate (o200k_base proxy)  $0.02      $0.0008  Goose, OpenCode, Kilo Code
3       84          51         GPT-6.1 Sol          84           estimate (o200k_base proxy)  $0.02      $0.0008  Amp
3       108         32         GPT-6.1 Sol          108          estimate (o200k_base proxy)  $0.02      $0.0011  Warp
2       282         19         GPT-6.1 Sol          282          estimate (o200k_base proxy)  $0.06      $0.0028  Factory Droid
4       303         51         GPT-6.1 Sol          303          estimate (o200k_base proxy)  $0.06      $0.0030  Augment Code
1       63          0          (no selected model)                                                                Google Jules
7       330         69         GPT-6.1 Sol          330          estimate (o200k_base proxy)  $0.07      $0.0033  Devin Local / Devin CLI

Advice (savings are estimates in o200k tokens per turn)
  HIGH duplicate CLAUDE.md, AGENTS.md [saves ~48 tokens/turn]
    3 identical lines are loaded twice every turn by cursor, copilot, factory-droid, augment, devin. Keep them in one file.
  HIGH contradiction AGENTS.md, CLAUDE.md [saves ~11 tokens/turn]
    Conflicting package manager: "Always use npm for installs and scripts in this repository." (AGENTS.md:3) vs "Always use pnpm for installs and scripts in this repository." (CLAUDE.md:5). (Removing the outdated line.)
  MEDIUM scopable-section CLAUDE.md:12 [saves ~128 tokens/turn]
    Section "API package" is mostly about packages/api/. Move it to packages/api/AGENTS.md (or CLAUDE.md), a path-scoped rule, or an instructions file ...

(Excerpt of the output for test/fixtures/mixed with --models gpt-6.1-sol.)

Related MCP server: contextburn

What it counts, and how

Claude's tokenizer is not public, and Anthropic's docs note that newer Claude models produce noticeably more tokens than older ones for the same text. turntax never shows an OpenAI count as a Claude or Gemini number: without a key those rows say "needs API key". The o200k column is a labelled reference count for every file.

Cost per N turns = tokens per turn × N × input price. The "cached" column assumes every turn reads the context from the prompt cache. Prices come from a dated table with a source link per model (turntax --list-models); pass --prices my-prices.json to use your own.

See docs/how-it-counts.md for the loading rules per agent and their sources.

Advice

All checks are deterministic. Each one shows an estimated saving in o200k tokens per turn.

• duplicate: identical lines in two files. A per-turn saving only when one tool loads both; otherwise flagged as drift risk.

• contradiction: conflicting rules reaching the same agent (package manager, indentation, quotes, semicolons).

• scopable-section: a section mostly about one directory that could load only there (nested file, path-scoped rule, applyTo).

• oversized-section: a large always-loaded section that could become an on-demand skill.

• size-limit: documented limits (CLAUDE.md under 200 lines, Codex's 32 KiB cut-off, Windsurf's 12,000 characters per rule and 6,000 for global rules, Cursor's 500 lines, Factory's 80,000 characters, Augment's 49,512 characters).

• misconfigured: files a tool ignores, such as .md files in .cursor/rules or Windsurf rules without a trigger.

• stale-reference: referenced files or npm run scripts that no longer exist.

• non-english: non-English sections. The token multiplier is measured only against an English version you supply (--english CLAUDE.md=CLAUDE.en.md) or, if you opt in, a Claude translation (--translate).

Fixing

npx turntax --fix            # print a diff of the safe fixes; nothing is changed
npx turntax --fix --write    # apply them

--fix handles four kinds of advice:

• duplicate: removes lines from a file when every agent that reads the file also reads the same lines every turn from another file. Skill, command and subagent files are left whole. A CLAUDE.md that repeats at least 80% of the lines of the AGENTS.md next to it, at least 3 of them full sentences, and every file it imports, keeps only its own lines and imports the rest with @AGENTS.md, as Claude Code's docs suggest; agents that read both stop paying for the copy, and the summary says how much any agent now reads on top.

• scopable-section: moves a section into the folder it is about: the same file name there (packages/api/CLAUDE.md), a nested AGENTS.md, or a Claude Code rule with paths:. The folder the section is about must already exist, and sections with @imports or relative links stay put. A section moves together with its subsections.

• oversized-section: moves a large procedure (a section of at least --section-tokens, 600 by default, whose heading names a release, deploy, setup or similar procedure, or that is mostly numbered steps; each subsection must pass too) into a new skill, .claude/skills/<name>/SKILL.md for Claude Code or .agents/skills/<name>/SKILL.md for Codex, and leaves the heading with a one-line pointer. Only the skill's name and description load every turn. Many agents now read one or both folders, but not all of them do (Zed reads only .agents/skills, for example), so by default this fix is often refused; --tools limits the check to the agents you use.

• misconfigured: renames .md files in .cursor/rules to .mdc.

Before a fix is kept, the files are classified again with the change applied: every agent must still read every line it read before, including files reached through @imports, and as often as before. The exceptions are moved text: a section moved into a folder may load only when the agent works there, and a procedure moved into a skill only when the skill is used. Otherwise the fix is skipped and the reason is printed. A moved section can reach agents that did not read it before (a nested AGENTS.md is read by more agents than a CLAUDE.md); the summary names them. By default that check covers every supported agent, so a section of AGENTS.md usually stays where it is, because some agents read only the root file. --tools codex,claude-code limits the check to the agents you use.

In preview the diff is on stdout and works with git apply, and the summary is on stderr; with --write the summary is on stdout. --write checks that no file changed since it was read and writes nothing if one did. Symlinked files, their targets and files in symlinked folders are never changed. Contradictions and stale references need a person to decide, so --fix leaves them alone. Files that are also documents for people (DESIGN.md, CONTEXT.md, agents.md, Agents.md, Claude.md) are never changed, and no file is created whose name differs from one of them only by case.

The web page shows the same fixes for any repository you load, with a downloadable fix.patch.

Config file

Put defaults in turntax.config.json in the audited folder (or pass --config <file>). Command-line flags override it.

{
  "tools": ["claude-code", "codex", "cursor"],
  "models": ["claude-sonnet-5.5", "gpt-6.1-sol"],
  "maxTokens": 8000,
  "maxIncrease": 500,
  "ignore": ["docs/agents/archive/", "**/*.draft.md"]
}

Keys: tools, models, turns, maxTokens, maxIncrease (applies only with --compare or --baseline), sectionTokens, and ignore: globs of instruction files to leave out, relative to the audited folder (* and ? within a folder name, ** across folders, a trailing / for a whole folder). Ignored files are also left out of --compare and are never changed by --fix. Unknown keys and ids are errors.

MCP servers

Every MCP server an agent starts adds its tool definitions (names, descriptions, input schemas) to the context. turntax reads the agents' MCP configs and lists the servers each agent starts:

A server name an agent finds in several configs is counted once, by the agent's documented precedence (or config order, marked as assumed); the other copies are listed as overridden. Tool filters in the configs (Codex enabled_tools/disabled_tools, Gemini includeTools/excludeTools, Factory enabledTools/disabledTools, Copilot tools, Kiro, Roo and Windsurf disabledTools) are applied.

Counting the tool definitions needs the servers running, so it is opt-in:

npx turntax --mcp

--mcp starts each stdio server with the command from its config (in the repository folder; environment variables are expanded into commands, URLs and headers the repository config defines, as the agents themselves do) and connects to streamable HTTP servers, asks for tools/list, and counts the definitions as JSON with o200k. A server several agents share is started once. Environment variables in the config (${VAR}, ${VAR:-default}, ${env:VAR}, {env:VAR}, ${workspaceFolder}) are expanded; servers that need a VS Code ${input:...} prompt, legacy SSE and WebSocket servers and servers that fail or time out (--mcp-timeout, default 30 s) are listed with the reason and not counted.

Claude Code defers MCP tool definitions by default (tool search): only the tool names are in context until a tool is used, so its row counts the names and shows the full size separately. A server with alwaysLoad: true, or a tool marked _meta["anthropic/alwaysLoad"], is counted in full. Server instructions load at the start of every session and are counted for Claude Code (cut at 4,096 characters, as Claude Code does), Gemini CLI, Codex and Kilo Code, whose docs say they reach the context.

MCP tokens are an o200k estimate and are not included in the model cost columns.

User-level files

Agents also read instruction files from your home directory in every project. --include-user adds them in a separate user/turn column and a total:

npx turntax --include-user

It reads only the locations the agents document (the full list, per agent, is in docs/how-it-counts.md): for example ~/.claude/, ~/.codex/AGENTS.md, ~/.agents/, ~/.gemini/, ~/.copilot/, ~/.cursor/skills, Windsurf's and Devin's global_rules.md and ~/.config/devin/, Cline's global rules, ~/.roo/, ~/.kiro/, ~/.junie/, ~/.config/goose/, ~/.config/amp/, ~/.config/opencode/, ~/.config/zed/AGENTS.md, ~/.factory/, ~/.augment/, ~/.config/kilo/, ~/.warp/ and ~/.trae/skills, plus the user MCP configs above. Contradictions and duplicates between user-level and repository files are reported. User-level files never count toward the badge or --max-tokens, so the result does not depend on who runs it.

Comparing with a baseline

npx turntax --compare origin/main            # instruction files at a git ref, read with git show
npx turntax --json --out before.json         # or save a report ...
npx turntax --baseline before.json           # ... and compare with it later
npx turntax --compare origin/main --max-increase 200   # exit 1 when any agent grows by more than 200 tokens

The comparison covers repository instruction files (o200k), per agent and per file, and lists MCP servers that were added or removed. It never changes the working tree.

GitHub Action

The repository is also an Action, listed on the GitHub Marketplace; npx turntax init writes this workflow and a turntax.config.json for you. On pull requests it compares with the base branch, posts one comment (updated on every push), writes the report to the job summary and marks findings on the lines they are about in the pull request's changed files (GitHub shows up to 10 warnings and 10 notices per step):

name: turntax
on: pull_request
permissions:
  contents: read
  pull-requests: write
jobs:
  turntax:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: TorutDev/turntax@v0.8.0
        with:
          max-increase: 500      # optional: fail when an agent grows by more than 500 tokens per turn
          # max-tokens: 8000     # optional: absolute budget
          # args: --tools claude-code,cursor
          # annotations: false   # optional: no findings on the changed lines

The comment opens with a table of the change per agent and per file. When --fix has safe fixes, it says how many and how many tokens they save, with the diff folded; the full report is folded below. A turntax.config.json in the repository is read as well. Pull requests from forks get a read-only token, so the comment step is skipped there; the job summary still has the report.

pre-commit

The repository is also a pre-commit hook, which fails a commit when an agent goes over a budget. Set the budget with args, or with maxTokens in turntax.config.json; without one the hook always passes. It runs npx turntax@<version>, so it needs Node.js 22 or later on your PATH, and it does not run on pre-commit.ci, which has no system hooks.

repos:
  - repo: https://github.com/TorutDev/turntax
    rev: v0.8.0
    hooks:
      - id: turntax
        args: [--max-tokens, '8000']

Options

turntax [path]
  --tools <ids>             agents to report (default: every agent that reads a file); see --help
  --models <ids>            see --list-models (default: claude-sonnet-5.5, gpt-6.1-sol, gemini-3.8-flash)
  --turns <n>               cost projection (default 100)
  --json | --md | --badge   output format (default: terminal table)
  --pr-comment              Markdown for a pull request comment (change first, fixes and full report folded)
  --annotations             GitHub Actions annotations, one per finding (--path-prefix <dir> for a subfolder)
  --out <file>              write the output to a file
  --max-tokens <n>          exit 1 when any tool loads more than n o200k tokens per turn (CI gate)
  --include-user            add user-level files and MCP servers from your home directory
  --mcp                     start the configured MCP servers and count their tool definitions
  --mcp-timeout <s>         seconds to wait for each MCP server (default 30)
  --baseline <file.json>    compare with an earlier --json report
  --compare <git-ref>       compare with the instruction files at a git ref
  --max-increase <n>        exit 1 when any tool grows by more than n tokens per turn vs the baseline
  --claude-api              exact Claude counts (ANTHROPIC_API_KEY)
  --gemini-api              Gemini counts (GEMINI_API_KEY)
  --english <file=en>       measure a language multiplier against your English version
  --translate               opt-in Claude translation for that measurement (needs --claude-api)
  --prices <file.json>      your own price table
  --section-tokens <n>      threshold for oversized sections (default 600)
  --fix                     print a diff of the safe fixes (--tools limits the agents whose context is kept)
  --write                   with --fix, apply the changes
  --config <file>           read defaults from this file (default: turntax.config.json, if any)

When a selected model's price was checked more than 90 days ago, turntax says so on stderr; check the source or pass --prices.

turntax init [path]         add .github/workflows/turntax.yml and turntax.config.json (--tools, --max-increase,
                            --force, --dry-run); existing files are kept unless --force
turntax mcp                 run turntax as an MCP server over stdio (see below)

Agents and MCP

turntax mcp (MCP Registry: io.github.TorutDev/turntax) is a read-only MCP server with two tools: audit (the report for a folder) and fix_preview (the --fix diff, never written). It reads turntax.config.json like the CLI and does not start MCP servers or call vendor APIs.

claude mcp add turntax -- npx -y turntax mcp

For Cursor and clients with an mcpServers map, add { "mcpServers": { "turntax": { "command": "npx", "args": ["-y", "turntax", "mcp"] } } } to their MCP config (VS Code's .vscode/mcp.json calls the map servers). On native Windows, run it through cmd /c npx -y turntax mcp if your client cannot start npx directly.

This repository is also a Claude Code plugin marketplace. The plugin adds the server and a /turntax:context-tax command:

/plugin marketplace add TorutDev/turntax
/plugin install turntax@turntax

VS Code

vscode/ is a VS Code extension on the same engine: the largest per-turn cost of the open folder in the status bar, a line above each instruction file with its tokens per turn and the agents that load it, findings in the Problems panel, the report as a Markdown preview, and safe fixes shown as diffs and written only when you choose Apply. It is offline and uses the folder's turntax.config.json; its one setting, turntax.codeLens, turns the line above each file off.

Install it from the VS Code Marketplace or, for Cursor, Windsurf and VSCodium, Open VSX, or run:

code --install-extension torut.turntax

Privacy

Offline by default: no network calls, no telemetry. File contents are sent only when you pass --claude-api, --gemini-api or --translate, and only to that vendor's API. Keys are read from environment variables and sent only in request headers.

--mcp runs the commands in the MCP configs it finds and connects to the configured HTTP servers, as the agents themselves would. Run it only on repositories whose configs you trust. turntax prints each command before starting it. --include-user reads files in your home directory and never sends them anywhere. --fix --write is the only option that changes files, and only repository files.

Web page

web/ is a static page that runs the same engine in the browser with the offline tokenizer only. Drop a repository folder or paste files; nothing is uploaded.

It can also load a public GitHub repository: it lists the files with the GitHub API (api.github.com, unauthenticated, 60 requests per hour) and downloads only the instruction files, and files they @import, from raw.githubusercontent.com. Those are the only network requests the page makes, and only when you load a repository or open a ?repo= link.

npm run build:web

Badge

npx turntax --badge

The badge shows the largest always-loaded total among the tools found, counted with o200k_base: repository files plus measured project MCP tools, never user-level files.

Limits

• Assumes the agent starts at the repository root. User-level files are counted only with --include-user; rules kept in an app's settings or a team dashboard (Cursor user rules, Copilot organization instructions) are not counted.

• Counts instruction files and, with --mcp, MCP tool definitions; not the agent's own system prompt or built-in tools.

• Loading rules change. Each rule cites the vendor page and the date it was checked; please open an issue when one is out of date.

Contributing

Loading rules and prices go out of date. Corrections with a link to the vendor's page are the most useful contribution; see CONTRIBUTING.md.

License

MIT